Deep learning and artificial intelligence, in general, is advancing scientific discovery and technological inventions through its ability to extract inherently hidden features and map it to output in a highly complex multi-dimensional space. Synthesis of electromagnetic (EM) structures with nearly arbitrary with desired functional properties is such an example of a high dimensional optimization space. In this article, we employ deep convolutional neural network (CNN) to allow robust and rapid prediction of scattering properties of nearly arbitrary planar electromagnetic structures on chip. Utilizing this, the work reports an mm-wave PA in 90-nm SiGe with a novel deep learning-enabled inverse design of low-loss, broadband output matching network that achieves a PAE of 16%-24.7%, a saturation power of 16.7-19.5 dBm across Psat, 3 dB bandwidth of 30-94 GHz (103.2%), while supporting both single-carrier high-speed modulation and concurrent multiband multi-Gb/s non-constant amplitude modulation. The Psat, 3 dB bandwidth covers from 5G band up to W-band and is higher than all reported mm-wave silicon PAs which have peak PAE > 20% and demonstrates for the first time concurrent multiband (triple-band) transmission with superior performance at multi-Gb/s.
All Science Journal Classification (ASJC) codes
- Condensed Matter Physics
- Electrical and Electronic Engineering
- Broadband PA
- Inverse design
- Machine learning